bbourn

Can you elaborate on what you mean by “it’s almost assumed..”? I’m actively working on datasheet parsing at scale at Zenode.ai, and I haven’t read anything specifically about this…

This is basically just a vent about all the sh!t i've dealt with over the years as a EE

Thank you! Yeah, missing pricing and availability is a tough one, but I'm really hoping we'll be able to get it from the distributors in return to funneling better traffic towards their sites...

We've actually heard this from a few industries, there's definitely a lot of different vertical AI opportunities here! Right now we're very focused on just doing electronic components, but maybe someday!

Yeah, this was actually a nightmare . We had to create "canonicals", which ended up being a mix of regex, machine learning and good old manual human labor. Fortunately, we could usually rely on the datasheet as the 'source of truth', but even that approach failed when faced with the overwhelming amount of part data we were shifting through.

That's the hope, but thus far we've been heads down building, and haven't had much time to chat with those folks.

From all of our customer interviews, almost everyone used Digikey and the occasional manufacturer website to find parts.

Because of that, we're seeing many people put the parameters they're looking for directly into the search. After a few searches, it starts to be clear that you can add more information directly into the search, so queries get a little more complex.

We're getting ready to add features to the AI to help you (interactively) navigate the range of available parts to narrow down your choices and understand what's available when you don't know where to look or what parameters to use. You should be able to ask 'I need to convert a parallel RGB display bus into MIPI' and the AI will show you different families of parts that could do that.

ChatGPT has definitely spoiled us all, but we're working to improve performance every day. Thus far we've usually prioritized accuracy over speed since there isn't much room for error in this industry, but hopefully our current response time is still faster than reading everything yourself. The sources not linking is a bug, should have it fixed today/tomorrow!

Well, that's super annoying, we definitely don't want that, thanks for finding and letting us know! We'll have it fixed shortly.

Digikey's parametrics were the best in industry prior to ours, but they were still really painful to use because of the strings. But to even get to the parametrics, you first have to know in advance the category and subcategory names of the part you're looking for. There are over 3000 different categories, so even once you're pretty good, you're always asking yourself "is what i'm looking for in the logic category or the amplifier category". With Zenode, all this is handled using natural language! And we did a huge overhaul on the parametrics, we turned the strings into structured values and built modern filters that are just waaay easier to use.

Octopart is a bit of a different beast from us. They're crushing the pricing and availability data, which is actually hard to get (a LOT of business development to get access to the realtime values from the distributors themselves). Plus they have a really massive catalog, definitely broader than ours, but they don't have the depth (i.e. the detailed spec and documentation data) that we do because we went to the manufacturer websites to obtain that.

We're learning a lot here, since this always happened AFTER we handed over the BOM. It certainly seems like a pain, particularly when a supplier ends up being a critical bottleneck (always late, overcharging, etc). The number one requested feature from these folks is a way to validate an alternate component, something they can use for themselves BEFORE passing it over to us design engineers for approval (heard some pretty spicy takes on doing this, apparently working with EE's isn't always sunshine and rainbows...)

Yeah, Moat's are tricky in the age of AI. We think there's always room for a standalone interface in this niche sector, given the sheer amount of data that needs to get presented (with no room for hallucinations). But beyond the data curation (which is actually quite hard) and the custom UI, there's also all the tuning data to optimize for the workflows that come next, like Alternates, schematic awareness, etc.

But yes, OpenAI continues to be the 800lb gorilla 'boogeyman' crushing startups with every new update. Our goal is to remain cockroaches until we find solid distribution!

Thanks, that's why we built it! We're big digikey fans, and have spent hundreds of hours on the site, but that doesn't mean it's without room for improvement!

Solid questions, you've clearly been burned before lol. 1) Document revisions can be a bit hard to handle, but we currently look for any dates or document numbers present, and when possible we just show the latest. However, plenty of times there are no dates OR revision numbers, and we'll get different documents from multiple sources. In this case we show all the documents, and have prompted the AI to flag this risk to the user if the question is answered differently between documents. Not the ideal approach, but working with what we're getting...

2) Hmm, not really. You can share the chat, and the specific response, but the only way to export the source would be to screenshot it... Adding it to our list of potential next features!

Awesome, that was one of the top reasons we built it, tired of clicking on strings lol. What are you working on specifically? Would love to chat and see what we can do to help!

No, we had those built a few versions back, but nobody used them and they were a pain to maintain so we took them out

This is actually well said. One of the issues we had to overcome with Zenode was the fact that people reference the same information in completely different ways. For example, "what's the nominal power draw" and "what's the typical power output" are the same query, but to an LLM, these are different.

We had to do a ton of work to get the tool to consistently return the same value (and still sometimes it shits the bed).

So yeah, even light technical jargon can confuse the heck out of an LLM, which is why that wasn't the only tool that's necessary to build this.

Try us out, would be curious to get your feedback! www.zenode.ai

Damn, yeah, we added the ability for the AI to set filters based on the input over the weekend and it has definitely slowed things down. We've got a rust script thats going out today to speed things up and add more functionality (like giving you the count of parts that each value in a filter is associated with), that should speed things up pretty dramatically....

Thanks for the try though, we'll definitely be improving this dramatically over the next few weeks!

AI for Electronics; www.zenode.ai

Hey, don't tell everyone our secret sauce! ;)

But yeah, you hit it on the head. LLMs alone were not enough to be able to read the datasheet. It took a mix of different algorithms, including OCR, Computer Vision, Neural Nets, and yes, LLMs to be able to consistently read all the differences between different manufactures, categories, etc.

Give us a whirl at www.zenode.ai and let me know what you think, we just launched our MVP and while there's a ton to improve we think it's pretty helpful!

We think it is, but filters are just the tip of the iceberg for what a true AI for electronics can do. Our service also uses Machine Learning to rank the results according what components you already think will or will not work for your service. It also looks at clusters within the results and pulls an example part from them, so that you don't have to look through a bunch of similar parts.

You should check it out, www.zenode.ai!

You should try www.Zenode.ai and tell us what you think! We did a lot of work to show all the components that could work, and use the AI to rank them according to what would work best for your needs. We still have a TON of things to do, but as hardware engineers ourselves we've found it useful for own projects!

This is exactly how and why we built www.Zenode.AI. It's a search engine for electronic components that uses and AI model trained to read the component datasheets

AI is the perfect tool for this task, but LLMs alone aren't capable of doing this. It took us multiple algorithms working together to read all the tables, equations, footnotes, etc. Now we've extracted the information from over 10M datasheets, and created a parts database that you can search across using natural language.

We're actually planning to hit Show Hacker News this week or next, just ironing out a few details first!

This is the kind of software that makes me excited, the future is going to be so much less wasted resources and so much more time spent building!

This is amazing, thank you for building!!!

Well that's just plain rude.

Fair point, we've actually been thinking along the same lines. What if the bottle just lit up for 20s? And had the proper steps on the label?

Hi Hacker News!

My gf is an ICU nurse and challenged me to build a device that helps people wash their hands properly (I wanted to build a ventilator, but she said that wasn't super helpful....:(

This is my first attempt, it's a soap dispenser that marks your hand with a stamp that fades with at least 20 seconds of hand washing—the minimum time recommended by the CDC.

Very open to comments both constructive and commending! :)

But.... Homie wants to play.... :_(

Love what you guys are doing, we’re doing something similar via the trash can, we should chat!!! Brandon[at]GarbiCan.com